Is this Real? Image Steganography without Embedding Spoofs Fake Image Detector

Steganography is a technology that conceals secret information in the carrier in an imperceptible way. Traditional embedding-based steganography are easily detected by steganalysis methods because of the inevitable modifications. Steganography-without-Embedding (SwE) methods benefit from the fact that they do not modify the carrier and can resist the traditional steganalysis which distinguishs stego images from cover images. However, in practice, SwE still face the challenge of fake image detectors. In this paper, we propose a SwE based on adversarial examples to improve steganographic security against both CS-steganalysis (distinguish cover and stego) and RF-steganalysis (distinguish real image and fake image). Two cases of RF-steganalysis, white-box and black-box, are considered. To improve the generator learning ability and the information recovery ability, we introduce the attention module in both the generator and extractor. Experimental results show that the proposed method has considerable extraction rate and could effectively resist both CS-steganalysis and RF-steganalysis.